Humanoid Robot Development

We trained a humanoid with 22-DoF dexterous hands to assemble model cars, operate syringes, sort poker cards, fold/roll shirts, all learned primarily from 20,000+ hours of egocentric human video with no robot in the loop. Humans are the most scalable embodiment on the planet. We discovered a near-perfect log-linear scaling law (R² = 0.998) between human video volume and action prediction loss, and this loss directly predicts real-robot success rate. Humanoid robots will be the end game, because they are the practical form factor with minimal embodiment gap from humans. Call it the Bitter Lesson of robot hardware: the kinematic similarity lets us simply retarget human finger motion onto dexterous robot hand joints. No learned embeddings, no fancy transfer algorithms needed. Relative wrist motion + retargeted 22-DoF finger actions serve as a unified action space that carries through from pre-training to robot execution. Our recipe is called "EgoScale": - Pre-train GR00T N1.5 on 20K hours of human video, mid-train with only 4 hours (!) of robot play data with Sharpa hands. 54% gains over training from scratch across 5 highly dexterous tasks. - Most surprising result: a *single* teleop demo is sufficient to learn a never-before-seen task. Our recipe enables extreme data efficiency. - Although we pre-train in 22-DoF hand joint space, the policy transfers to a Unitree G1 with 7-DoF tri-finger hands. 30%+ gains over training on G1 data alone. The scalable path to robot dexterity was never more robots. It was always us. - Website: https://lnkd.in/gxzgeP-2 - Paper: https://lnkd.in/g7PJdz_8

In just ONE year, humanoid robots at the CCTV Spring Festival Gala went from “cool machines” to something that felt… human. What do you think? 2025 → 2026. The difference? Not incremental. Exponential. What changed in 12 months? 📊 The Data Behind the Leap: • AI model capability has been doubling at unprecedented rates (training compute for frontier models has grown >10x in short cycles). • Latency in edge AI systems is now measured in single-digit milliseconds — enabling real-time motion response. • Actuator precision and torque density in humanoid robotics improved significantly, enabling smoother micro-movements. • Multimodal AI (vision + audio + spatial awareness) accuracy has crossed 90%+ benchmarks in controlled environments. • Reinforcement learning in simulation can now compress “years” of physical training into weeks. Result? 2025: Pre-programmed choreography. 2026: Real-time adaptive interaction. We are witnessing the shift from: 🔹 Robots as automation to 🔹 Robots as embodied AI platforms And here’s the bigger implication: When physical AI converges with high-performance edge compute, robotics stops being hardware-centric… and becomes software-defined. The real revolution isn’t the robot you saw on stage. It’s the AI stack running inside it. If this is the progress visible in public within 12 months, imagine what’s happening inside R&D labs right now. Humanoids are no longer a science experiment. They are becoming infrastructure. 2026 is the year robotics started to feel personal. #AI #Robotics #PhysicalAI #Humanoids #DeepLearning #EdgeAI #Innovation

A major breakthrough in reinforcement learning for robot training and the NeurIPS 2025 Best Paper. When training robots to walk, navigate, or manipulate objects, RL researchers have usually been using relatively shallow networks—typically 2-5 layer MLPs mapping sensor readings to motor commands. Attempts to go deeper have failed because training becomes unstable and performance degrades. Prior work attributed these failures to RL's sparse feedback: you might get one bit of information after thousands of decisions, so the ratio of signal to parameters is tiny. In this paper, the authors show that the problem was architectural rather than fundamental. With residual connections, layer normalization, and Swish activation—surprisingly standard elsewhere but not in control RL—you can train networks up to 1000+ layers. The paper demonstrates that gains from adding layers aren't gradual: at certain depth thresholds, agents acquire new behaviors. A simulated humanoid learns to walk upright only at 16 layers; at 256 layers, it learns to vault over walls. Read online and ask questions when blocked: https://lnkd.in/e7jzcc5G Download the PDF: https://lnkd.in/e2Bk8pdU The full list of the most important AI paper of 2025: https://lnkd.in/ekfaXgwJ

Nature already solved some of the hardest parts. Strength and softness at the same time. Control without fragile complexity. Safety without sacrificing capability. Copying those patterns can skip years of trial and error in a lab. Allonic is building a robot hand using “3D tissue braiding,” basically weaving high strength fibers around a minimal rigid skeleton the way connective tissue wraps around bone. Instead of hundreds of screws, bearings, cables, and fiddly joints, it’s one continuous process that forms the tendons, soft tissue, and compliant structure together. The result looks more like biology than engineering. Once robot bodies become easier to manufacture and easier to redesign, the upgrade loop becomes constant. Faster iteration, lower costs, better dexterity, more safety, then repeat. Pretty cool. Follow Endrit Restelica for more.

NVIDIA has trained humanoid robots to move like Cristiano Ronaldo, LeBron James, and Kobe Bryant, using neural networks on real hardware in their GEAR lab. And unlike typical robotics demos (that tend to speed up footage) here the video has been slowed down to show just how fluid the movement is. Their latest innovation, "ASAP", is a "real2sim2real"model, meaning it first captures the athletes' movements from video footage before creating a 3D digital model from the data and then finally recreating this in the real world via the humanoid robot. The last part of this process is the most difficult however (coining a new phrase "easier simmed than done") as the real world introduces countless unexpected physical variables that can screw up the final result. To overcome this "sim2real gap", NVIDIA uses a "neural net" (see explanatory video in the comments) to predict and adjust for these physical complexities in real time (just as our own brains do). This hybrid simulation approach combines classical physics with AI to create spookily human looking robotic motion. NVIDIA is open-sourcing this work—so you can learn more and even gain access to their clever physics models on their project page (I've included the link in the comments below). So, what do you think? Could robot sports-bots end up joining (or even competing against) human teams in the future?

A few years ago, what you’re seeing here would’ve required a motion capture studio, expensive rigs, specialised suits, and a budget most companies wouldn’t even consider touching. Today, it starts with curiosity, smart experimentation, and a team willing to break things. This week, a few of our Enfectors were experimenting with hand tracking and motion controls, mapping real human movement onto an AI character we’ve been developing. What you get is a hyper-realistic digital character that moves, reacts, and behaves like a real person, not an animated approximation. Under the hood, this touches a fascinating stack: • Real-time hand and body tracking • Motion retargeting • Skeletal rigs and inverse kinematics • AI-assisted character generation • Real-time rendering and animation pipelines What excites me most isn’t the tech itself, but what it unlocks. This kind of experimentation is helping us expand how we think about: • AI brand characters and spokespeople • Virtual influencers and digital ambassadors • Product explainers and immersive storytelling • Training, demos, and interactive brand experiences And yes, all of this is being built by a team in Sri Lanka, for brands anywhere in the world.

All valuable work will increasingly be done by Human-AI hybrids. An insightful research paper identifies both challenges and good practices from multiple case studies to propose an overall framework. The authors propose that generating effective human-AI hybrids is divided into two phases: Construction - in which Technical implementers design the architecture of the hybrid - and Execution - where Organizational implementers facilitate how participants engage and interact. They suggest 3 primary success factors: 🔧 Interface and Technical Design focuses on making AI systems accessible and reliable through code-free interfaces. The technical architecture should allow rapid testing of different approaches while being supported by effective data curation strategies. 🧠 Human Capability Development prepares people to work effectively with AI systems through training, in critical assessment and prompting techniques. Employees must understand AI's capabilities and limitations, and develop skills to integrate AI into existing workflows. 🤝 The Collaboration Framework structures successful human-AI interaction through aligned mental models and clear role definitions. It emphasizes improving underperforming areas rather than disrupting successful processes, while ensuring both human and AI agents contribute their unique strengths to achieve optimal outcomes.

The Humanoid Robot Race Has Started — But the Real Innovation Is the AI Infrastructure Behind BMW’s deployment of Figure 03, developed by Figure AI, at its Spartanburg plant is an important milestone in the evolution of industrial automation. Figure is building general-purpose humanoid robots designed to operate in environments created for humans. Unlike traditional industrial robots that are typically fixed in place and programmed for specific tasks, humanoid robots are designed to navigate existing workplaces, use human tools, and adapt to changing conditions through AI. And Figure is not alone. The humanoid robotics market is becoming increasingly competitive, with companies such as Tesla with Optimus, Agility Robotics with Digit, and other emerging players investing heavily in bringing general-purpose robots into real-world environments. BMW spent nearly a year testing Figure 02, where the robot supported body shop operations and contributed to the production of more than 30,000 BMW X3 vehicles. That was a controlled environment: fixed parts, predictable positions, and repeatable movements. With Figure 03, BMW is moving into a more complex challenge: logistics sequencing in Hall 52, supporting the flow of parts for vehicle assembly, including the BMW X3 and upcoming electrified models. This is a much harder problem. In logistics, parts may arrive in different positions or orientations. The robot must understand its surroundings, identify objects, decide how to handle them, and adapt its movements in real time. That is the difference between traditional automation and Physical AI. But the real breakthrough is not just the robot. A humanoid robot only creates value when connected to a broader digital ecosystem: • Digital twins that simulate factories and optimize workflows before deployment. • Integrated IT/OT infrastructure connecting robots, production systems, and operations. • Low-latency networks and edge computing to process vision, touch, and movement data in real time. • AI-powered quality systems that continuously monitor operations and provide rapid feedback. The impact on people may be just as important as the technology. As robots take on repetitive and physically demanding tasks, employees can increasingly focus on higher-value work: supervising intelligent systems, solving exceptions, improving processes, and making decisions where human judgment matters most. The lesson for organizations is clear: Physical AI will not scale through hardware alone. It requires the right digital foundation — and a workforce prepared for a new way of working. The robot is the visible innovation. The real transformation is the combination of AI, infrastructure, and human capability. https://lnkd.in/eDUPt8_r

Tesla isn’t building a robot. They’re rebuilding the human hand. Tesla Optimus (Gen 3) When Elon Musk says the hardest part isn’t AI, balance, or autonomy—but the hands—that tells you everything. Human hands: ~27–30 degrees of freedom. Tendon-driven. Muscles mostly in the forearm. Ridiculous force control. Replicating that? Savage engineering. Here’s what most people miss: 1. Dexterity = control bandwidth. Not strength. You need ultra-low latency actuation, torque density in tiny volumes, minimal backlash, compliance, and thermal stability at duty cycle. That’s a controls + hardware problem. 2. The supply chain doesn’t exist. So you vertically integrate. Motors. Gearboxes. Inverters. Controllers. Same EV strategy. New battlefield. 3. Tendon routing is the cheat code. Biology uses remote actuation. Lightweight. Compact. Elegant. But hard to model and even harder to scale. Expect heavy use of: – Closed-loop torque sensing – Predictive grasp modeling – Self-calibration – Learning-based manipulation AI meets mechatronics. 4. Why Gen 3 matters If it nails tool use, soft-object handling, cable routing, and two-hand coordination… Humanoids stop being demos. They become labor. And that changes manufacturing, logistics, eldercare, retail—everything. Big idea: Autonomy without dexterity is a Roomba. Dexterity + autonomy is a workforce. The next industrial revolution won’t be software-first. It’ll be torque-density-first. Five fingers wide. SDVGuru.com #Tesla #Optimus #HumanoidRobots #RoboticsEngineering #AI #EmbodiedAI #Mechatronics #DeepTech #Automation #FutureOfWork #Actuators #AdvancedManufacturing #VerticalIntegration

Closing the sim-to-real gap in humanoid robotics requires massive simulation throughput and high-fidelity physics validation. WPP recently detailed their engineering pipeline, showing how they reduced reinforcement learning cycle times for complex humanoid locomotion from 24 hours down to less than 60 minutes. The hardware architecture relies on Google Cloud’s new G4 VMs (powered by NVIDIA RTX PRO 6000 Blackwell GPUs) running NVIDIA Isaac Sim, integrated closely with DeepMind’s MuJoCo physics engine. The mechanics: The team mapped raw human mocap data (over 200 degrees of freedom) down to a constrained 29-DOF OpenUSD digital twin. By leveraging a P2P GPU topology to bypass central processing bottlenecks, the infrastructure executed over 3 billion simulations in under an hour. The virtual environment continuously introduced physical micro-variances—simulated pushes, shifting floor friction, and momentum changes—to train the model against the chaos of the real world. The resulting reinforcement learning model was condensed into a highly efficient ONNX policy and deployed directly to the physical robot. This edge policy processes live IMU and joint telemetry to output immediate, stabilized motor commands. Reaching this scale of simulation volume is the precise engineering mechanism that allows control policies to handle unstructured physical deployment. To support the research, Unitree has open-sourced the underlying RL code on GitHub. Blog post : https://lnkd.in/g4-gWzTP #Robotics #PhysicalAI #ReinforcementLearning #MuJoCo #GoogleCloud #IsaacSim #Engineering